reviews at, well, the end of the month. This time, there was a short delay of about a week.
The reason was conference travel.
To put things into perspective for readers unfamiliar with the conference game in machine learning (ML) research, here is the story, slightly abridged.
If you do ML research, you usually want to publish your work as a paper at a conference. Interesting but unpublished work does not count for much, so it is often publish or perish. To publish a paper, however, you first need a fitting outlet. In ML, the main outlets are three premier conferences: ICML, NeurIPS, and ICLR.
The International Conference on Machine Learning (ICML) was first held in 1980. For its first decade, it was organized as a workshop; afterwards, it continued as a full conference.
In 1987, the first Conference on Neural Information Processing Systems (NeurIPS) was held in Denver, USA.
ICLR joined the ranks more recently. In 2013, researchers founded the International Conference on Learning Representations (ICLR).
To publish a paper at one of these conferences, you have to go through the sometimes annoying process of peer review. First, you submit an anonymized version of your paper, meaning that you remove names, affiliations, and other identifying information. This version is then anonymously reviewed by — hopefully — more knowledgeable peers from your research field.
You receive their feedback, get some opportunity to clarify or improve the paper, and then wait for another month or two. Based on the quality of the initial submission and how well you addressed the reviewers’ concerns, your paper will either be accepted or rejected.
This is the important part: acceptance rates are low. Usually, only somewhere between 20% and 30% of submitted papers are accepted. It is a competitive game.
Thus, if you manage to publish a paper at one of these conferences, you can consider yourself lucky. You can then visit the conference itself, which brings me back to the topic of this article.
Getting to the conference
The conferences are held annually at different locations across the world. Depending on where you live, you might have quite some distance to cover.
If you are based in China and the conference takes place in South Korea, you can get there relatively quickly. If you are coming from Europe or the United States, however, you will probably spend most of a day travelling. The same applies in reverse when a conference is held in the States.
My last visit to such a large conference was ICLR 2024, which was hosted in Vienna. For me, that was quite comfortable to reach.
Last month — and this explains the delay in writing this monthly review — I attended ICML in South Korea. The conference itself was nice, it was the getting there took quite some time.
And it is not only the travelling that matters. There is also the organizational overhead and, perhaps more importantly, the opportunity cost.
Let me explain.
The hidden overhead of conference travel
First, you need to book the flights and the hotel. Usually, you also need to stay within certain per-diem limits, meaning that you have a daily budget of a certain amount to spend on accommodation.
You therefore survey the hotels close to the conference venue, check whether they fit the budget, get one approved, and finally book it. The same applies to the flights. Altogether, this can easily take half a day — sometimes a full one.
Then the actual travel begins…
Getting to the airport takes time. Waiting at the airport takes more time. The flight itself can take more than ten hours. Finally, travelling from the destination airport to the hotel adds another hour or two.
After roughly 24 hours, you have reached your destination. Great.
But then there is the jet lag. The conference takes place in another time zone, so your body needs time to adjust — ideally while you are already supposed to attend talks, meet people, and present your work.
On the way back, the same thing happens in reverse. You spend another day travelling home, followed by another round of jet lag.
All in all, you have now spent two or three days organizing and completing the journey. On top of that come the actual conference days, which might add another three to seven days.
In total, the conference can occupy around ten days.
That would be fine if the rest of life simply paused during that time. Unfortunately, it does not.
The backlog waits at home
When you return, your regular work is still there. Emails have accumulated. Tasks have been waiting. Projects have continued without you — or, perhaps worse, have not continued at all.
Which is what happened to me. After returning, I first had to work through my personal backlog. There were things to tend to, tasks to redo, and several loose ends to pick up again. On top of that, I had to prepare a journal revision and a conference rebuttal.
The conference itself had ended, but the disruption around it had not.
And that, to me, is the real downside of conference travel. It is not merely the long flight or the jet lag. It is this interruption of your normal rhythm: the preparation beforehand, the travel itself, the time spent at the event, and the backlog waiting afterwards.
All this easily kills a week or two of progress on a new research project. And, a week or two more on a paper can make quite a large difference. It could be another proof, it could be coding a difficult experiment, writing a complete draft. For all these things, you need extended, uninterrupted time, ideally for several days in a row. Deep work, it is (a favorite topic of mine, see also my recent review of my last 8.5 years in machine learning),
This is not to say that conferences are unnecessary. No, conferences are valuable. You meet other researchers, discuss ideas, see new work, and finally put faces to names you may have known only from papers. Still, that value comes at a considerable cost in time and energy.
This month’s lesson was therefore rather simple: when planning conference travel, do not count only the days written in the conference program. Count the preparation, the journey, the recovery, and the work that will be waiting when you return.
Otherwise, a five-day conference can quietly consume two full weeks without taking into account the opportunity costs.
Instagram Engineering
Shopify Engineering